diff --git a/app/api/search/route.ts b/app/api/search/route.ts
new file mode 100644
index 000000000..fa17b7b9c
--- /dev/null
+++ b/app/api/search/route.ts
@@ -0,0 +1,63 @@
+import { NextRequest, NextResponse } from "next/server";
+
+import { searchStyleSlugs } from "@/lib/retrieval/style-search-service";
+import {
+  checkRateLimit,
+  createRateLimitHeaders,
+  getRequestClientKey,
+} from "@/lib/security/rate-limit";
+
+const MAX_QUERY_LENGTH = 200;
+
+// Rankings change only with a deployment; identical queries can be shared.
+const SEARCH_CACHE_CONTROL =
+  "public, max-age=300, s-maxage=3600, stale-while-revalidate=86400";
+
+/**
+ * Hybrid style search: BM25 + vector + RRF over the curated catalog.
+ *
+ * Returns ranked slugs only. Callers already hold the catalog (from
+ * `/api/styles`) and join on slug, so this stays small and cacheable. `mode`
+ * says whether the vector path answered or the keyword path served alone.
+ */
+export async function GET(request: NextRequest) {
+  const rateLimit = checkRateLimit({
+    namespace: "style-search",
+    key: getRequestClientKey(request),
+    limit: 60,
+    windowMs: 60 * 1000,
+  });
+  if (!rateLimit.allowed) {
+    return NextResponse.json(
+      { error: "Too many requests. Try again later." },
+      { status: 429, headers: createRateLimitHeaders(rateLimit) },
+    );
+  }
+
+  const query = (request.nextUrl.searchParams.get("q") ?? "").trim();
+  if (!query) {
+    return NextResponse.json({ error: "Missing query parameter q." }, { status: 400 });
+  }
+  if (query.length > MAX_QUERY_LENGTH) {
+    return NextResponse.json(
+      { error: `Query must be at most ${MAX_QUERY_LENGTH} characters.` },
+      { status: 400 },
+    );
+  }
+
+  try {
+    const result = await searchStyleSlugs(query);
+    return NextResponse.json(
+      { query, total: result.results.length, ...result },
+      // A degraded answer is cached briefly so the vector path gets retried soon.
+      {
+        headers: {
+          "Cache-Control": result.mode === "hybrid" ? SEARCH_CACHE_CONTROL : "public, max-age=60",
+        },
+      },
+    );
+  } catch (error) {
+    console.error("[api/search] failed:", error);
+    return NextResponse.json({ error: "Search is temporarily unavailable." }, { status: 503 });
+  }
+}
diff --git a/lib/retrieval/style-search-service.ts b/lib/retrieval/style-search-service.ts
new file mode 100644
index 000000000..9dba083b5
--- /dev/null
+++ b/lib/retrieval/style-search-service.ts
@@ -0,0 +1,134 @@
+/**
+ * @module retrieval/style-search-service
+ *
+ * The production entry point for hybrid style search (BM25 + vector + RRF).
+ *
+ * `hybrid-search.ts` is the algorithm; this module is the wiring a server
+ * process needs around it: one corpus index per process, the vector index read
+ * from disk, and the embedding provider built from the environment.
+ *
+ * Every missing piece degrades rather than fails. No `DASHSCOPE_API_KEY`, no
+ * `.data/style-vectors.json`, an index built for another model, or a slow
+ * embedding call all fall back to the keyword path, and the response says
+ * which one answered. That keeps search available on a host that has not been
+ * given the key yet.
+ */
+
+import { chunkStyles } from "@/lib/retrieval/chunk-styles";
+import {
+  DEFAULT_EMBEDDING_DIMENSIONS,
+  DEFAULT_EMBEDDING_MODEL,
+  createDashScopeEmbeddingProvider,
+  type EmbeddingProvider,
+} from "@/lib/retrieval/embedding";
+import {
+  createHybridSearcher,
+  type DegradeReason,
+  type HybridSearcher,
+} from "@/lib/retrieval/hybrid-search";
+import { JsonFileVectorStore, type VectorStore } from "@/lib/retrieval/vector-store";
+import { styles } from "@/lib/styles/registry";
+
+/** A query answered from the network should not wait on a slow embedding. */
+const QUERY_EMBEDDING_TIMEOUT_MS = 1_500;
+const QUERY_CACHE_LIMIT = 500;
+
+export type StyleSearchMode = "hybrid" | "keyword";
+
+export interface StyleSearchResult {
+  mode: StyleSearchMode;
+  /** Set when the vector path was unavailable for this query. */
+  degradeReason?: DegradeReason;
+  results: Array<{ slug: string; score: number }>;
+}
+
+interface ServiceState {
+  searcher: HybridSearcher;
+  /** Why the vector path is off for every query, when it is. */
+  staticDegrade?: DegradeReason;
+}
+
+let statePromise: Promise<ServiceState> | null = null;
+const queryCache = new Map<string, StyleSearchResult>();
+
+function createProvider(): EmbeddingProvider | null {
+  try {
+    return createDashScopeEmbeddingProvider({ timeoutMs: QUERY_EMBEDDING_TIMEOUT_MS, maxAttempts: 1 });
+  } catch {
+    return null;
+  }
+}
+
+async function openVectorStore(): Promise<VectorStore | null> {
+  try {
+    const store = await JsonFileVectorStore.open({
+      dimensions: DEFAULT_EMBEDDING_DIMENSIONS,
+      model: DEFAULT_EMBEDDING_MODEL,
+    });
+    return store.size() > 0 ? store : null;
+  } catch {
+    // A stale index (other model or dimensions) is unusable, not fatal.
+    return null;
+  }
+}
+
+async function buildState(): Promise<ServiceState> {
+  const provider = createProvider();
+  const vectorStore = await openVectorStore();
+  const searcher = createHybridSearcher({
+    chunks: chunkStyles(styles),
+    embeddingProvider: vectorStore ? provider : null,
+    vectorStore,
+    topK: styles.length,
+    vectorTimeoutMs: QUERY_EMBEDDING_TIMEOUT_MS,
+  });
+  const staticDegrade: DegradeReason | undefined = !provider ? "no-provider" : !vectorStore ? "no-store" : undefined;
+  return { searcher, staticDegrade };
+}
+
+function getState(): Promise<ServiceState> {
+  if (!statePromise) {
+    statePromise = buildState().catch((error: unknown) => {
+      statePromise = null;
+      throw error;
+    });
+  }
+  return statePromise;
+}
+
+export async function searchStyleSlugs(query: string): Promise<StyleSearchResult> {
+  const key = query.trim().toLowerCase();
+  const cached = queryCache.get(key);
+  if (cached) return cached;
+
+  const { searcher, staticDegrade } = await getState();
+  let degradeReason: DegradeReason | undefined = staticDegrade;
+  const hits = await searcher.search(key, {
+    onDegrade: (info) => {
+      degradeReason ??= info.reason;
+    },
+  });
+
+  const result: StyleSearchResult = {
+    mode: degradeReason ? "keyword" : "hybrid",
+    ...(degradeReason ? { degradeReason } : {}),
+    results: hits.map((hit) => ({ slug: hit.slug, score: hit.score })),
+  };
+
+  // Only full-quality answers are cached, so a transient embedding failure
+  // does not pin a keyword-only ranking for that query.
+  if (result.mode === "hybrid" || staticDegrade) {
+    if (queryCache.size >= QUERY_CACHE_LIMIT) {
+      const oldest = queryCache.keys().next().value;
+      if (oldest !== undefined) queryCache.delete(oldest);
+    }
+    queryCache.set(key, result);
+  }
+  return result;
+}
+
+/** Exposed for tests. */
+export function resetStyleSearchService(): void {
+  statePromise = null;
+  queryCache.clear();
+}
diff --git a/packages/core/src/discovery/index.ts b/packages/core/src/discovery/index.ts
index 0cfa82bf5..85691dae1 100644
--- a/packages/core/src/discovery/index.ts
+++ b/packages/core/src/discovery/index.ts
@@ -45,4 +45,4 @@ export {
   clearRemoteCache,
 } from "./remote";
 
-export type { DataOrigin, Sourced, RemoteOptions } from "./remote";
+export type { DataOrigin, Sourced, RemoteOptions, SearchRanking } from "./remote";
diff --git a/packages/core/src/discovery/remote.ts b/packages/core/src/discovery/remote.ts
index ac47e2e80..a8a4dc53c 100644
--- a/packages/core/src/discovery/remote.ts
+++ b/packages/core/src/discovery/remote.ts
@@ -13,10 +13,11 @@
  * correct by construction; keeping the bundle as a fallback means losing the
  * network degrades to stale rather than to broken.
  *
- * Ranking stays local. The API supplies which styles exist; the bundled
- * scoring decides how they rank. Moving scoring server-side would create two
- * implementations of the same logic, and they would drift the way the data
- * just did.
+ * Query ranking prefers the site's hybrid search (`/api/search`: BM25 +
+ * vector + RRF). Its vector path needs an embedding key and a prebuilt index,
+ * neither of which can ship inside an npm package, so it has to run
+ * server-side. The bundled scorer stays as the fallback when that endpoint is
+ * unreachable, so a query always gets an answer.
  */
 
 import {
@@ -44,8 +45,15 @@ export interface Sourced<T> {
   readonly origin: DataOrigin;
   /** Why the live catalogue was not used, when it was not. */
   readonly fallbackReason?: string;
+  /**
+   * Which ranker ordered a query's results: the site's hybrid search, its
+   * keyword-only degradation, or the bundled scorer.
+   */
+  readonly ranking?: SearchRanking;
 }
 
+export type SearchRanking = "hybrid" | "keyword" | "local";
+
 export interface RemoteOptions {
   /** Override for testing or self-hosting. */
   readonly baseUrl?: string;
@@ -111,6 +119,7 @@ async function fetchJsonUncached<T>(
   cacheKey: string,
   options: RemoteOptions,
   generation: number,
+  tripCircuit: boolean,
 ): Promise<FetchResult<T>> {
   const timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
   const controller = new AbortController();
@@ -122,7 +131,7 @@ async function fetchJsonUncached<T>(
       headers: { accept: "application/json" },
     });
     if (!response.ok) {
-      openCircuit(baseUrl, generation);
+      if (tripCircuit) openCircuit(baseUrl, generation);
       return { error: `HTTP ${response.status}` };
     }
 
@@ -136,7 +145,7 @@ async function fetchJsonUncached<T>(
     }
     return { value };
   } catch (error) {
-    openCircuit(baseUrl, generation);
+    if (tripCircuit) openCircuit(baseUrl, generation);
     const message = error instanceof Error ? error.message : String(error);
     return {
       error: /abort/i.test(message) ? `timed out after ${timeoutMs}ms` : message,
@@ -146,7 +155,15 @@ async function fetchJsonUncached<T>(
   }
 }
 
-async function fetchJson<T>(path: string, options: RemoteOptions): Promise<FetchResult<T>> {
+/**
+ * `tripCircuit: false` is for optional endpoints: a site that predates one
+ * answers 404, and that must not take the catalogue offline with it.
+ */
+async function fetchJson<T>(
+  path: string,
+  options: RemoteOptions,
+  tripCircuit = true,
+): Promise<FetchResult<T>> {
   if (options.live === false) return { error: "live fetching disabled" };
 
   const baseUrl = normalizeBaseUrl(options.baseUrl ?? STYLEKIT_SITE_URL);
@@ -163,7 +180,7 @@ async function fetchJson<T>(path: string, options: RemoteOptions): Promise<Fetch
   if (existing) return (await existing) as FetchResult<T>;
 
   const generation = cacheGeneration;
-  const request = fetchJsonUncached<T>(path, baseUrl, cacheKey, options, generation);
+  const request = fetchJsonUncached<T>(path, baseUrl, cacheKey, options, generation, tripCircuit);
   inFlight.set(cacheKey, request as Promise<FetchResult<unknown>>);
   try {
     return await request;
@@ -340,11 +357,53 @@ export async function searchStylesLive(
 ): Promise<Sourced<{ total: number; results: StyleSummary[] }>> {
   const catalogue = await liveCatalogue(options);
   if ("error" in catalogue) {
-    return { data: searchWithPool(opts), origin: "bundled", fallbackReason: catalogue.error };
+    return {
+      data: searchWithPool(opts),
+      origin: "bundled",
+      fallbackReason: catalogue.error,
+      ...(opts.query?.trim() ? { ranking: "local" as const } : {}),
+    };
+  }
+
+  const query = opts.query?.trim();
+  if (query) {
+    const ranked = await rankedSlugs(query, options);
+    if (ranked && ranked.slugs.length > 0) {
+      const bySlug = new Map(catalogue.styles.map((style) => [style.slug, style]));
+      const pool = ranked.slugs
+        .map((slug) => bySlug.get(slug))
+        .filter((style): style is DesignStyle => style !== undefined);
+      // No query here: the pool is already in ranked order, and searchWithPool
+      // keeps that order while applying the category filter and limit.
+      return {
+        data: searchWithPool({ ...opts, query: undefined }, pool),
+        origin: "live",
+        ranking: ranked.mode,
+      };
+    }
+    return { data: searchWithPool(opts, catalogue.styles), origin: "live", ranking: "local" };
   }
   return { data: searchWithPool(opts, catalogue.styles), origin: "live" };
 }
 
+async function rankedSlugs(
+  query: string,
+  options: RemoteOptions,
+): Promise<{ slugs: string[]; mode: "hybrid" | "keyword" } | null> {
+  const response = await fetchJson<unknown>(
+    `/api/search?q=${encodeURIComponent(query)}`,
+    options,
+    false,
+  );
+  if ("error" in response || !isRecord(response.value)) return null;
+  const { mode, results } = response.value;
+  if ((mode !== "hybrid" && mode !== "keyword") || !Array.isArray(results)) return null;
+  const slugs = results
+    .map((result) => (isRecord(result) ? result.slug : undefined))
+    .filter((slug): slug is string => typeof slug === "string" && slug.length > 0);
+  return { slugs, mode };
+}
+
 export async function getStyleDetailLive(
   slug: string,
   options: RemoteOptions = {},
diff --git a/packages/mcp/src/tools.ts b/packages/mcp/src/tools.ts
index ce19e8a14..8e9f32ddf 100644
--- a/packages/mcp/src/tools.ts
+++ b/packages/mcp/src/tools.ts
@@ -30,6 +30,12 @@ const READ_ONLY = {
 
 const CATEGORIES = ["modern", "retro", "minimal", "expressive"] as const;
 
+const RANKING_LABEL = {
+  hybrid: "hybrid search (BM25 + vector + RRF)",
+  keyword: "keyword search (vector path unavailable)",
+  local: "bundled scorer (live search unavailable)",
+} as const;
+
 // Shared output shapes (so clients get typed structuredContent).
 const SUMMARY_SHAPE = {
   slug: z.string(),
@@ -209,6 +215,7 @@ Examples:
         count: z.number(),
         offset: z.number(),
         has_more: z.boolean(),
+        ranking: z.enum(["hybrid", "keyword", "local"]).optional(),
         results: z.array(z.object(SUMMARY_SHAPE)),
       },
       annotations: READ_ONLY,
@@ -228,7 +235,7 @@ Examples:
       const hasMore = offset + page.length < total;
       const lines = [
         `# StyleKit styles${query ? ` matching "${query}"` : ""}`,
-        `Found ${total} (showing ${page.length}${offset ? ` from offset ${offset}` : ""}).`,
+        `Found ${total} (showing ${page.length}${offset ? ` from offset ${offset}` : ""})${search.ranking ? ` · ranked by ${RANKING_LABEL[search.ranking]}` : ""}.`,
         "",
         ...page.map(
           (r) =>
@@ -241,6 +248,7 @@ Examples:
         count: page.length,
         offset,
         has_more: hasMore,
+        ...(search.ranking ? { ranking: search.ranking } : {}),
         results: page,
       });
     },
diff --git a/tests/unit/packages-core/remote-discovery.test.ts b/tests/unit/packages-core/remote-discovery.test.ts
index 8d901decc..b1b0982f8 100644
--- a/tests/unit/packages-core/remote-discovery.test.ts
+++ b/tests/unit/packages-core/remote-discovery.test.ts
@@ -235,4 +235,54 @@ describe("remote discovery", () => {
     await searchStylesLive({}, { baseUrl: "https://ttl.test", cacheTtlMs: 100 });
     expect(fetchMock).toHaveBeenCalledTimes(3);
   });
+
+  it("orders query results by the site's hybrid search and keeps the category filter", async () => {
+    const fetchMock = vi.fn((url: string) => {
+      if (url.includes("/api/search")) {
+        return Promise.resolve(
+          responseFor(url, {
+            mode: "hybrid",
+            results: [
+              { slug: "gamma", score: 3 },
+              { slug: "retro-one", score: 2 },
+              { slug: "alpha", score: 1 },
+              { slug: "unknown-slug", score: 0.5 },
+            ],
+          }),
+        );
+      }
+      return Promise.resolve(
+        catalogue([style("alpha"), style("gamma"), style("retro-one", "retro")]),
+      );
+    });
+    vi.stubGlobal("fetch", fetchMock);
+
+    const result = await searchStylesLive(
+      { query: "玻璃质感", category: "modern" },
+      { baseUrl: "https://hybrid.test" },
+    );
+    expect(result.origin).toBe("live");
+    expect(result.ranking).toBe("hybrid");
+    expect(result.data.results.map((r) => r.slug)).toEqual(["gamma", "alpha"]);
+    expect(fetchMock.mock.calls.some(([url]) => String(url).includes("q=%E7%8E%BB%E7%92%83"))).toBe(true);
+  });
+
+  it("falls back to the bundled scorer when search is missing, without backing off the catalogue", async () => {
+    const fetchMock = vi.fn((url: string) =>
+      Promise.resolve(
+        url.includes("/api/search")
+          ? responseFor(url, { error: "not found" }, 404)
+          : catalogue([style("alpha"), style("gamma")]),
+      ),
+    );
+    vi.stubGlobal("fetch", fetchMock);
+
+    const first = await searchStylesLive({ query: "gamma" }, { baseUrl: "https://old-site.test" });
+    expect(first.origin).toBe("live");
+    expect(first.ranking).toBe("local");
+    expect(first.data.results.map((r) => r.slug)).toEqual(["gamma"]);
+
+    const detail = await getStyleDetailLive("not-bundled-slug", { baseUrl: "https://old-site.test" });
+    expect(detail.fallbackReason).not.toBe("live source unreachable, backing off");
+  });
 });
diff --git a/tests/unit/style-search-service.test.ts b/tests/unit/style-search-service.test.ts
new file mode 100644
index 000000000..37dbacffb
--- /dev/null
+++ b/tests/unit/style-search-service.test.ts
@@ -0,0 +1,27 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+import { resetStyleSearchService, searchStyleSlugs } from "@/lib/retrieval/style-search-service";
+
+afterEach(() => {
+  resetStyleSearchService();
+  vi.unstubAllEnvs();
+});
+
+describe("style search service", () => {
+  it("serves keyword results when no embedding key is configured", async () => {
+    vi.stubEnv("DASHSCOPE_API_KEY", "");
+
+    const result = await searchStyleSlugs("glassmorphism");
+    expect(result.mode).toBe("keyword");
+    expect(result.degradeReason).toBe("no-provider");
+    expect(result.results.slice(0, 3).map((hit) => hit.slug)).toContain("glassmorphism");
+  });
+
+  it("finds styles for a colloquial Chinese query", async () => {
+    vi.stubEnv("DASHSCOPE_API_KEY", "");
+
+    const result = await searchStyleSlugs("毛玻璃");
+    expect(result.results.length).toBeGreaterThan(0);
+    expect(result.results.slice(0, 5).map((hit) => hit.slug)).toContain("glassmorphism");
+  });
+});